EXIN BCS Gen AI Award flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

Generative AI

A class of machine learning models that produce new content such as text, images or code by learning the statistical distribution of training data, rather than only classifying or ranking existing inputs.

Transformer

A neural network architecture built on self-attention, allowing each token to weigh all others in the input. It is the basis of most modern large language models.

Token

The atomic unit of text a language model processes, roughly three to four English characters. Context limits and pricing are usually expressed in tokens.

Context window

The maximum number of tokens a model can consider at once, covering both the prompt and the generated output. Exceeding it causes truncation or lost information.

Temperature

A sampling parameter that scales the model's output probabilities. Low values make responses focused and repeatable; high values increase diversity and randomness.

Top-p sampling

Nucleus sampling that keeps only the smallest set of tokens whose cumulative probability reaches a threshold p, then samples from that set.

Embedding

A dense numeric vector representing the meaning of text. Similar meanings produce nearby vectors, enabling semantic search and retrieval for augmented generation.

Hallucination

Fluent model output that is factually wrong or unsupported by any source. It occurs because the model predicts plausible text rather than verifying truth.

Knowledge cutoff

The date after which a model has no training data. Questions about later events cannot be answered reliably without retrieval or external tools.

Sycophancy

The tendency of language models to agree with a user's stated view or flatter the questioner, which can bias answers toward whatever position the prompt implies.

Lost in the middle

An observed weakness where models recall information placed at the start or end of a long context better than material buried in the middle.

Multimodal model

A model that accepts more than one input type, such as text plus images or audio. It can describe images but may misread charts and small text.

Data leakage risk

The exposure that occurs when confidential input is pasted into a third-party tool that may retain it or use it for training, breaching policy or regulation.

Prompt components

The standard building blocks of an effective prompt: role, task, context, input data, constraints and output format. Missing format specification is a common failure.

Few-shot prompting

Including a small number of worked examples in the prompt so the model imitates their pattern. Examples must be consistent, because errors in them are copied.

Chain-of-thought prompting

Asking the model to reason through intermediate steps before answering. It improves multi-step tasks but can be unnecessary on models that reason internally.

Retrieval-augmented generation

A pattern that retrieves relevant documents and inserts them into the prompt so the answer is grounded in supplied sources, reducing hallucination and enabling citations.

Prompt injection

An attack where untrusted text inside a prompt or retrieved document contains instructions that override the developer's intent, potentially triggering unsafe actions or data disclosure.

Output schema

A specified response structure, such as JSON with named fields or a fixed table, that makes model output machine-parseable and easier to validate automatically.

Prompt decomposition

Splitting a complex task into a sequence of smaller prompts, checking each intermediate result. It usually outperforms a single large prompt for multi-stage work.

System prompt

Instructions set by the application developer that define the model's role, tone and boundaries. They take priority over user turns but are not a security boundary.

EU AI Act risk tiers

A four-level classification of AI systems: unacceptable (prohibited), high-risk (strict obligations), limited-risk (transparency duties) and minimal-risk (no added duties).

GDPR and prompts

Personal data entered into prompts remains subject to GDPR, so a lawful basis, data minimisation and erasure rights apply to model inputs and stored outputs.

AI output copyright

In several jurisdictions, including the US, a work generated purely by AI without meaningful human authorship may not qualify for copyright protection.

Training data poisoning

A security attack that corrupts training data so the model learns biased or backdoored behaviour, which is hard to detect after deployment.

Membership inference

A privacy attack that determines whether a specific record was in a model's training data, potentially revealing sensitive information about individuals.

Human oversight

A governance control requiring a competent person to review or approve consequential AI-assisted decisions. Accountability remains with the deploying organisation.

Synthetic media labelling

The practice of marking AI-generated images, audio or video as synthetic, required in some jurisdictions to prevent deception and preserve public trust.

AI use-case screening

A filter for viable projects: high volume, language-heavy, review-tolerant, low error cost and evaluable against ground truth. Irreversible regulated decisions are excluded.

Build versus buy

A sourcing decision: hosted APIs offer speed and low fixed cost, while fine-tuned or self-hosted open models suit data residency, latency or high-volume unit costs.

Gen AI ROI estimate

Value equals time saved times loaded hourly cost times adoption rate, minus inference, integration and review costs. Adoption rate is usually the largest uncertainty.

Golden test set

A curated set of representative inputs with known correct outputs, used to measure model accuracy, faithfulness and hallucination rate before and after deployment.